def _build_graph(self, inputs):
        input, output = inputs
        input, output = input / 128.0 - 1, output / 128.0 - 1

        with argscope([Conv2D, Deconv2D],
                      W_init=tf.truncated_normal_initializer(stddev=0.02)), \
                argscope(LeakyReLU, alpha=0.2):
            with tf.variable_scope('gen'):
                fake_output = self.generator(input)
            with tf.variable_scope('discrim'):
                real_pred = self.discriminator(input, output)
            with tf.variable_scope('discrim', reuse=True):
                fake_pred = self.discriminator(input, fake_output)

        self.build_losses(real_pred, fake_pred)
        errL1 = tf.reduce_mean(tf.abs(fake_output - output), name='L1_loss')
        self.g_loss = tf.add(self.g_loss, LAMBDA * errL1, name='total_g_loss')
        add_moving_summary(errL1, self.g_loss)

        # tensorboard visualization
        if IN_CH == 1:
            input = tf.image.grayscale_to_rgb(input)
        if OUT_CH == 1:
            output = tf.image.grayscale_to_rgb(output)
            fake_output = tf.image.grayscale_to_rgb(fake_output)
        viz = (tf.concat([input, output, fake_output], 2) + 1.0) * 128.0
        viz = tf.cast(tf.clip_by_value(viz, 0, 255), tf.uint8, name='viz')
        tf.summary.image('input,output,fake', viz, max_outputs=max(30, BATCH))

        self.collect_variables()

    def build_graph(self, input, output):
        input, output = input / 128.0 - 1, output / 128.0 - 1

        with argscope([Conv2D, Conv2DTranspose], kernel_initializer=tf.truncated_normal_initializer(stddev=0.02)):
            with tf.variable_scope('gen'):
                fake_output = self.generator(input)
            with tf.variable_scope('discrim'):
                real_pred = self.discriminator(input, output)
                fake_pred = self.discriminator(input, fake_output)

        self.build_losses(real_pred, fake_pred)
        errL1 = tf.reduce_mean(tf.abs(fake_output - output), name='L1_loss')
        self.g_loss = tf.add(self.g_loss, LAMBDA * errL1, name='total_g_loss')
        add_moving_summary(errL1, self.g_loss)

        # tensorboard visualization
        if IN_CH == 1:
            input = tf.image.grayscale_to_rgb(input)
        if OUT_CH == 1:
            output = tf.image.grayscale_to_rgb(output)
            fake_output = tf.image.grayscale_to_rgb(fake_output)

        visualize_tensors('input,output,fake', [input, output, fake_output], max_outputs=max(30, BATCH))

        self.collect_variables()

    def _build_graph(self, inputs):
        input, nextinput = inputs

        cell = rnn.MultiRNNCell([rnn.LSTMBlockCell(num_units=param.rnn_size)
                                for _ in range(param.num_rnn_layer)])

        def get_v(n):
            ret = tf.get_variable(n + '_unused', [param.batch_size, param.rnn_size],
                                  trainable=False,
                                  initializer=tf.constant_initializer())
            ret = tf.placeholder_with_default(ret, shape=[None, param.rnn_size], name=n)
            return ret
        initial = (rnn.LSTMStateTuple(get_v('c0'), get_v('h0')),
                   rnn.LSTMStateTuple(get_v('c1'), get_v('h1')))

        embeddingW = tf.get_variable('embedding', [param.vocab_size, param.rnn_size])
        input_feature = tf.nn.embedding_lookup(embeddingW, input)  # B x seqlen x rnnsize

        input_list = tf.unstack(input_feature, axis=1)  # seqlen x (Bxrnnsize)

        outputs, last_state = rnn.static_rnn(cell, input_list, initial, scope='rnnlm')
        last_state = tf.identity(last_state, 'last_state')

        # seqlen x (Bxrnnsize)
        output = tf.reshape(tf.concat(outputs, 1), [-1, param.rnn_size])  # (Bxseqlen) x rnnsize
        logits = FullyConnected('fc', output, param.vocab_size, nl=tf.identity)
        tf.nn.softmax(logits / param.softmax_temprature, name='prob')

        xent_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
            logits=logits, labels=tf.reshape(nextinput, [-1]))
        self.cost = tf.reduce_mean(xent_loss, name='cost')
        summary.add_param_summary(('.*/W', ['histogram']))   # monitor histogram of all W
        summary.add_moving_summary(self.cost)

    def _build_graph(self, inputs, is_training):
        state, action, reward, next_state, isOver = inputs
        self.predict_value = self._get_DQN_prediction(state, is_training)
        action_onehot = tf.one_hot(action, NUM_ACTIONS)
        pred_action_value = tf.reduce_sum(self.predict_value * action_onehot, 1)    #N,
        max_pred_reward = tf.reduce_mean(tf.reduce_max(
            self.predict_value, 1), name='predict_reward')
        add_moving_summary(max_pred_reward)
        self.greedy_choice = tf.argmax(self.predict_value, 1)   # N,

        with tf.variable_scope('target'):
            targetQ_predict_value = self._get_DQN_prediction(next_state, False)    # NxA

            # DQN
            #best_v = tf.reduce_max(targetQ_predict_value, 1)    # N,

            # Double-DQN
            predict_onehot = tf.one_hot(self.greedy_choice, NUM_ACTIONS, 1.0, 0.0)
            best_v = tf.reduce_sum(targetQ_predict_value * predict_onehot, 1)

            target = reward + (1.0 - tf.cast(isOver, tf.float32)) * GAMMA * tf.stop_gradient(best_v)

        sqrcost = tf.square(target - pred_action_value)
        abscost = tf.abs(target - pred_action_value)    # robust error func
        cost = tf.select(abscost < 1, sqrcost, abscost)
        summary.add_param_summary([('conv.*/W', ['histogram', 'rms']),
                                   ('fc.*/W', ['histogram', 'rms']) ])   # monitor all W
        self.cost = tf.reduce_mean(cost, name='cost')

def multilevel_rpn_losses(
        multilevel_anchors, multilevel_label_logits, multilevel_box_logits):
    """
    Args:
        multilevel_anchors: #lvl RPNAnchors
        multilevel_label_logits: #lvl tensors of shape HxWxA
        multilevel_box_logits: #lvl tensors of shape HxWxAx4

    Returns:
        label_loss, box_loss
    """
    num_lvl = len(cfg.FPN.ANCHOR_STRIDES)
    assert len(multilevel_anchors) == num_lvl
    assert len(multilevel_label_logits) == num_lvl
    assert len(multilevel_box_logits) == num_lvl

    losses = []
    with tf.name_scope('rpn_losses'):
        for lvl in range(num_lvl):
            anchors = multilevel_anchors[lvl]
            label_loss, box_loss = rpn_losses(
                anchors.gt_labels, anchors.encoded_gt_boxes(),
                multilevel_label_logits[lvl], multilevel_box_logits[lvl],
                name_scope='level{}'.format(lvl + 2))
            losses.extend([label_loss, box_loss])

        total_label_loss = tf.add_n(losses[::2], name='label_loss')
        total_box_loss = tf.add_n(losses[1::2], name='box_loss')
        add_moving_summary(total_label_loss, total_box_loss)
    return total_label_loss, total_box_loss

def fpn_map_rois_to_levels(boxes):
    """
    Assign boxes to level 2~5.

    Args:
        boxes (nx4):

    Returns:
        [tf.Tensor]: 4 tensors for level 2-5. Each tensor is a vector of indices of boxes in its level.
        [tf.Tensor]: 4 tensors, the gathered boxes in each level.

    Be careful that the returned tensor could be empty.
    """
    sqrtarea = tf.sqrt(tf_area(boxes))
    level = tf.to_int32(tf.floor(
        4 + tf.log(sqrtarea * (1. / 224) + 1e-6) * (1.0 / np.log(2))))

    # RoI levels range from 2~5 (not 6)
    level_ids = [
        tf.where(level <= 2),
        tf.where(tf.equal(level, 3)),   # == is not supported
        tf.where(tf.equal(level, 4)),
        tf.where(level >= 5)]
    level_ids = [tf.reshape(x, [-1], name='roi_level{}_id'.format(i + 2))
                 for i, x in enumerate(level_ids)]
    num_in_levels = [tf.size(x, name='num_roi_level{}'.format(i + 2))
                     for i, x in enumerate(level_ids)]
    add_moving_summary(*num_in_levels)

    level_boxes = [tf.gather(boxes, ids) for ids in level_ids]
    return level_ids, level_boxes

    def build_graph(self, image, label):
        xys = np.array([(y, x, 1) for y in range(WARP_TARGET_SIZE)
                        for x in range(WARP_TARGET_SIZE)], dtype='float32')
        xys = tf.constant(xys, dtype=tf.float32, name='xys')    # p x 3

        image = image / 255.0 - 0.5  # bhw2

        def get_stn(image):
            stn = (LinearWrap(image)
                   .AvgPooling('downsample', 2)
                   .Conv2D('conv0', 20, 5, padding='VALID')
                   .MaxPooling('pool0', 2)
                   .Conv2D('conv1', 20, 5, padding='VALID')
                   .FullyConnected('fc1', 32)
                   .FullyConnected('fct', 6, activation=tf.identity,
                                   kernel_initializer=tf.constant_initializer(),
                                   bias_initializer=tf.constant_initializer([1, 0, HALF_DIFF, 0, 1, HALF_DIFF]))())
            # output 6 parameters for affine transformation
            stn = tf.reshape(stn, [-1, 2, 3], name='affine')  # bx2x3
            stn = tf.reshape(tf.transpose(stn, [2, 0, 1]), [3, -1])  # 3 x (bx2)
            coor = tf.reshape(tf.matmul(xys, stn),
                              [WARP_TARGET_SIZE, WARP_TARGET_SIZE, -1, 2])
            coor = tf.transpose(coor, [2, 0, 1, 3], 'sampled_coords')  # b h w 2
            sampled = BilinearSample('warp', [image, coor], borderMode='constant')
            return sampled

        with argscope([Conv2D, FullyConnected], activation=tf.nn.relu):
            with tf.variable_scope('STN1'):
                sampled1 = get_stn(image)
            with tf.variable_scope('STN2'):
                sampled2 = get_stn(image)

        # For visualization in tensorboard
        with tf.name_scope('visualization'):
            padded1 = tf.pad(sampled1, [[0, 0], [HALF_DIFF, HALF_DIFF], [HALF_DIFF, HALF_DIFF], [0, 0]])
            padded2 = tf.pad(sampled2, [[0, 0], [HALF_DIFF, HALF_DIFF], [HALF_DIFF, HALF_DIFF], [0, 0]])
            img_orig = tf.concat([image[:, :, :, 0], image[:, :, :, 1]], 1)  # b x 2h  x w
            transform1 = tf.concat([padded1[:, :, :, 0], padded1[:, :, :, 1]], 1)
            transform2 = tf.concat([padded2[:, :, :, 0], padded2[:, :, :, 1]], 1)
            stacked = tf.concat([img_orig, transform1, transform2], 2, 'viz')
            tf.summary.image('visualize',
                             tf.expand_dims(stacked, -1), max_outputs=30)

        sampled = tf.concat([sampled1, sampled2], 3, 'sampled_concat')
        logits = (LinearWrap(sampled)
                  .FullyConnected('fc1', 256, activation=tf.nn.relu)
                  .FullyConnected('fc2', 128, activation=tf.nn.relu)
                  .FullyConnected('fct', 19, activation=tf.identity)())
        tf.nn.softmax(logits, name='prob')

        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
        cost = tf.reduce_mean(cost, name='cross_entropy_loss')

        wrong = tf.to_float(tf.logical_not(tf.nn.in_top_k(logits, label, 1)), name='incorrect_vector')
        summary.add_moving_summary(tf.reduce_mean(wrong, name='train_error'))

        wd_cost = tf.multiply(1e-5, regularize_cost('fc.*/W', tf.nn.l2_loss),
                              name='regularize_loss')
        summary.add_moving_summary(cost, wd_cost)
        return tf.add_n([wd_cost, cost], name='cost')

    def build_losses(self, logits_real, logits_fake):
        """D and G play two-player minimax game with value function V(G,D)

          min_G max _D V(D, G) = IE_{x ~ p_data} [log D(x)] + IE_{z ~ p_fake} [log (1 - D(G(z)))]

        Args:
            logits_real (tf.Tensor): discrim logits from real samples
            logits_fake (tf.Tensor): discrim logits from fake samples produced by generator
        """
        with tf.name_scope("GAN_loss"):
            score_real = tf.sigmoid(logits_real)
            score_fake = tf.sigmoid(logits_fake)
            tf.summary.histogram('score-real', score_real)
            tf.summary.histogram('score-fake', score_fake)

            with tf.name_scope("discrim"):
                d_loss_pos = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
                    logits=logits_real, labels=tf.ones_like(logits_real)), name='loss_real')
                d_loss_neg = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
                    logits=logits_fake, labels=tf.zeros_like(logits_fake)), name='loss_fake')

                d_pos_acc = tf.reduce_mean(tf.cast(score_real > 0.5, tf.float32), name='accuracy_real')
                d_neg_acc = tf.reduce_mean(tf.cast(score_fake < 0.5, tf.float32), name='accuracy_fake')

                d_accuracy = tf.add(.5 * d_pos_acc, .5 * d_neg_acc, name='accuracy')
                self.d_loss = tf.add(.5 * d_loss_pos, .5 * d_loss_neg, name='loss')

            with tf.name_scope("gen"):
                self.g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
                    logits=logits_fake, labels=tf.ones_like(logits_fake)), name='loss')
                g_accuracy = tf.reduce_mean(tf.cast(score_fake > 0.5, tf.float32), name='accuracy')

            add_moving_summary(self.g_loss, self.d_loss, d_accuracy, g_accuracy)

    def build_graph(self, image_pos):
        image_pos = image_pos / 128.0 - 1

        z = tf.random_normal([self.batch, self.zdim], name='z_train')
        z = tf.placeholder_with_default(z, [None, self.zdim], name='z')

        with argscope([Conv2D, Conv2DTranspose, FullyConnected],
                      kernel_initializer=tf.truncated_normal_initializer(stddev=0.02)):
            with tf.variable_scope('gen'):
                image_gen = self.generator(z)
            tf.summary.image('generated-samples', image_gen, max_outputs=30)

            alpha = tf.random_uniform(shape=[self.batch, 1, 1, 1],
                                      minval=0., maxval=1., name='alpha')
            interp = image_pos + alpha * (image_gen - image_pos)

            with tf.variable_scope('discrim'):
                vecpos = self.discriminator(image_pos)
                vecneg = self.discriminator(image_gen)
                vec_interp = self.discriminator(interp)

        # the Wasserstein-GAN losses
        self.d_loss = tf.reduce_mean(vecneg - vecpos, name='d_loss')
        self.g_loss = tf.negative(tf.reduce_mean(vecneg), name='g_loss')

        # the gradient penalty loss
        gradients = tf.gradients(vec_interp, [interp])[0]
        gradients = tf.sqrt(tf.reduce_sum(tf.square(gradients), [1, 2, 3]))
        gradients_rms = symbolic_functions.rms(gradients, 'gradient_rms')
        gradient_penalty = tf.reduce_mean(tf.square(gradients - 1), name='gradient_penalty')
        add_moving_summary(self.d_loss, self.g_loss, gradient_penalty, gradients_rms)

        self.d_loss = tf.add(self.d_loss, 10 * gradient_penalty)

        self.collect_variables()

        def LSGAN_losses(real, fake):
            d_real = tf.reduce_mean(tf.squared_difference(real, 1), name='d_real')
            d_fake = tf.reduce_mean(tf.square(fake), name='d_fake')
            d_loss = tf.multiply(d_real + d_fake, 0.5, name='d_loss')

            g_loss = tf.reduce_mean(tf.squared_difference(fake, 1), name='g_loss')
            add_moving_summary(g_loss, d_loss)
            return g_loss, d_loss

def rpn_losses(anchor_labels, anchor_boxes, label_logits, box_logits):
    """
    Args:
        anchor_labels: fHxfWxNA
        anchor_boxes: fHxfWxNAx4, encoded
        label_logits:  fHxfWxNA
        box_logits: fHxfWxNAx4

    Returns:
        label_loss, box_loss
    """
    with tf.device('/cpu:0'):
        valid_mask = tf.stop_gradient(tf.not_equal(anchor_labels, -1))
        pos_mask = tf.stop_gradient(tf.equal(anchor_labels, 1))
        nr_valid = tf.stop_gradient(tf.count_nonzero(valid_mask, dtype=tf.int32), name='num_valid_anchor')
        nr_pos = tf.count_nonzero(pos_mask, dtype=tf.int32, name='num_pos_anchor')

        valid_anchor_labels = tf.boolean_mask(anchor_labels, valid_mask)
    valid_label_logits = tf.boolean_mask(label_logits, valid_mask)

    with tf.name_scope('label_metrics'):
        valid_label_prob = tf.nn.sigmoid(valid_label_logits)
        summaries = []
        with tf.device('/cpu:0'):
            for th in [0.5, 0.2, 0.1]:
                valid_prediction = tf.cast(valid_label_prob > th, tf.int32)
                nr_pos_prediction = tf.reduce_sum(valid_prediction, name='num_pos_prediction')
                pos_prediction_corr = tf.count_nonzero(
                    tf.logical_and(
                        valid_label_prob > th,
                        tf.equal(valid_prediction, valid_anchor_labels)),
                    dtype=tf.int32)
                summaries.append(tf.truediv(
                    pos_prediction_corr,
                    nr_pos, name='recall_th{}'.format(th)))
                precision = tf.to_float(tf.truediv(pos_prediction_corr, nr_pos_prediction))
                precision = tf.where(tf.equal(nr_pos_prediction, 0), 0.0, precision, name='precision_th{}'.format(th))
                summaries.append(precision)
        add_moving_summary(*summaries)

    label_loss = tf.nn.sigmoid_cross_entropy_with_logits(
        labels=tf.to_float(valid_anchor_labels), logits=valid_label_logits)
    label_loss = tf.reduce_mean(label_loss, name='label_loss')

    pos_anchor_boxes = tf.boolean_mask(anchor_boxes, pos_mask)
    pos_box_logits = tf.boolean_mask(box_logits, pos_mask)
    delta = 1.0 / 9
    box_loss = tf.losses.huber_loss(
        pos_anchor_boxes, pos_box_logits, delta=delta,
        reduction=tf.losses.Reduction.SUM) / delta
    box_loss = tf.div(
        box_loss,
        tf.cast(nr_valid, tf.float32), name='box_loss')

    add_moving_summary(label_loss, box_loss, nr_valid, nr_pos)
    return label_loss, box_loss

    def _build_graph(self, inputs):
        x, y, label = inputs
        x, y = self.embed([x, y])

        with tf.variable_scope(tf.get_variable_scope(), reuse=True):
            tf.identity(self.embed(inputs[0]), name="emb")

        cost = symbf.siamese_cosine_loss(x, y, label, scope="loss")
        self.cost = tf.identity(cost, name="cost")
        add_moving_summary(self.cost)

 def get_feature_match_loss(self, feats_real, feats_fake):
     losses = []
     for real, fake in zip(feats_real, feats_fake):
         loss = tf.reduce_mean(tf.squared_difference(
             tf.reduce_mean(real, 0),
             tf.reduce_mean(fake, 0)),
             name='mse_feat_' + real.op.name)
         losses.append(loss)
     ret = tf.add_n(losses, name='feature_match_loss')
     add_moving_summary(ret)
     return ret

    def build_graph(self, x, y, label):
        single_input = x
        x, y = self.embed([x, y])

        with tf.variable_scope(tf.get_variable_scope(), reuse=True):
            tf.identity(self.embed(single_input), name="emb")

        cost = siamese_cosine_loss(x, y, label, scope="loss")
        cost = tf.identity(cost, name="cost")
        add_moving_summary(cost)
        return cost

    def _build_graph(self, inputs):
        """This function should build the model which takes the input variables
        and define self.cost at the end"""

        # inputs contains a list of input variables defined above
        image, label = inputs

        # In tensorflow, inputs to convolution function are assumed to be
        # NHWC. Add a single channel here.
        image = tf.expand_dims(image, 3)

        image = image * 2 - 1   # center the pixels values at zero

        # The context manager `argscope` sets the default option for all the layers under
        # this context. Here we use 32 channel convolution with shape 3x3
        with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu, out_channel=32):
            logits = (LinearWrap(image)
                      .Conv2D('conv0')
                      .MaxPooling('pool0', 2)
                      .Conv2D('conv1')
                      .Conv2D('conv2')
                      .MaxPooling('pool1', 2)
                      .Conv2D('conv3')
                      .FullyConnected('fc0', 512, nl=tf.nn.relu)
                      .Dropout('dropout', 0.5)
                      .FullyConnected('fc1', out_dim=10, nl=tf.identity)())

        tf.nn.softmax(logits, name='prob')   # a Bx10 with probabilities

        # a vector of length B with loss of each sample
        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
        cost = tf.reduce_mean(cost, name='cross_entropy_loss')  # the average cross-entropy loss

        correct = tf.cast(tf.nn.in_top_k(logits, label, 1), tf.float32, name='correct')
        accuracy = tf.reduce_mean(correct, name='accuracy')

        # This will monitor training error (in a moving_average fashion):
        # 1. write the value to tensosrboard
        # 2. write the value to stat.json
        # 3. print the value after each epoch
        train_error = tf.reduce_mean(1 - correct, name='train_error')
        summary.add_moving_summary(train_error, accuracy)

        # Use a regex to find parameters to apply weight decay.
        # Here we apply a weight decay on all W (weight matrix) of all fc layers
        wd_cost = tf.multiply(1e-5,
                              regularize_cost('fc.*/W', tf.nn.l2_loss),
                              name='regularize_loss')
        self.cost = tf.add_n([wd_cost, cost], name='total_cost')
        summary.add_moving_summary(cost, wd_cost, self.cost)

        # monitor histogram of all weight (of conv and fc layers) in tensorboard
        summary.add_param_summary(('.*/W', ['histogram', 'rms']))

    def build_graph(self, image, label):
        """This function should build the model which takes the input variables
        and return cost at the end"""

        # In tensorflow, inputs to convolution function are assumed to be
        # NHWC. Add a single channel here.
        image = tf.expand_dims(image, 3)

        image = image * 2 - 1   # center the pixels values at zero

        # The context manager `argscope` sets the default option for all the layers under
        # this context. Here we use 32 channel convolution with shape 3x3
        with argscope([tf.layers.conv2d], padding='same', activation=tf.nn.relu):
            l = tf.layers.conv2d(image, 32, 3, name='conv0')
            l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
            l = tf.layers.conv2d(l, 32, 3, name='conv1')
            l = tf.layers.conv2d(l, 32, 3, name='conv2')
            l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
            l = tf.layers.conv2d(l, 32, 3, name='conv3')
            l = tf.layers.flatten(l)
            l = tf.layers.dense(l, 512, activation=tf.nn.relu, name='fc0')
            l = tf.layers.dropout(l, rate=0.5,
                                  training=get_current_tower_context().is_training)
        logits = tf.layers.dense(l, 10, activation=tf.identity, name='fc1')

        # a vector of length B with loss of each sample
        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
        cost = tf.reduce_mean(cost, name='cross_entropy_loss')  # the average cross-entropy loss

        correct = tf.cast(tf.nn.in_top_k(logits, label, 1), tf.float32, name='correct')
        accuracy = tf.reduce_mean(correct, name='accuracy')

        # This will monitor training error & accuracy (in a moving average fashion). The value will be automatically
        # 1. written to tensosrboard
        # 2. written to stat.json
        # 3. printed after each epoch
        train_error = tf.reduce_mean(1 - correct, name='train_error')
        summary.add_moving_summary(train_error, accuracy)

        # Use a regex to find parameters to apply weight decay.
        # Here we apply a weight decay on all W (weight matrix) of all fc layers
        # If you don't like regex, you can certainly define the cost in any other methods.
        wd_cost = tf.multiply(1e-5,
                              regularize_cost('fc.*/kernel', tf.nn.l2_loss),
                              name='regularize_loss')
        total_cost = tf.add_n([wd_cost, cost], name='total_cost')
        summary.add_moving_summary(cost, wd_cost, total_cost)

        # monitor histogram of all weight (of conv and fc layers) in tensorboard
        summary.add_param_summary(('.*/kernel', ['histogram', 'rms']))
        # the function should return the total cost to be optimized
        return total_cost

    def _build_graph(self, inputs):
        image, label = inputs
        image = ImageNetModel.image_preprocess(image, bgr=self.image_bgr)
        if self.data_format == 'NCHW':
            image = tf.transpose(image, [0, 3, 1, 2])

        logits = self.get_logits(image)
        loss = ImageNetModel.compute_loss_and_error(logits, label)
        wd_loss = regularize_cost(self.weight_decay_pattern,
                                  tf.contrib.layers.l2_regularizer(self.weight_decay),
                                  name='l2_regularize_loss')
        add_moving_summary(loss, wd_loss)
        self.cost = tf.add_n([loss, wd_loss], name='cost')

    def _build_graph(self, inputs):
        """This function should build the model which takes the input variables
        and define self.cost at the end"""

        # inputs contains a list of input variables defined above
        image, label = inputs

        # In tensorflow, inputs to convolution function are assumed to be
        # NHWC. Add a single channel here.
        image = tf.expand_dims(image, 3)

        image = image * 2 - 1   # center the pixels values at zero

        l = tf.layers.conv2d(image, 32, 3, padding='same', activation=tf.nn.relu, name='conv0')
        l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
        l = tf.layers.conv2d(l, 32, 3, padding='same', activation=tf.nn.relu, name='conv1')
        l = tf.layers.conv2d(l, 32, 3, padding='same', activation=tf.nn.relu, name='conv2')
        l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
        l = tf.layers.conv2d(l, 32, 3, padding='same', activation=tf.nn.relu, name='conv3')
        l = tf.layers.flatten(l)
        l = tf.layers.dense(l, 512, activation=tf.nn.relu, name='fc0')
        l = tf.layers.dropout(l, rate=0.5,
                              training=get_current_tower_context().is_training)
        logits = tf.layers.dense(l, 10, activation=tf.identity, name='fc1')

        tf.nn.softmax(logits, name='prob')   # a Bx10 with probabilities

        # a vector of length B with loss of each sample
        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
        cost = tf.reduce_mean(cost, name='cross_entropy_loss')  # the average cross-entropy loss

        correct = tf.cast(tf.nn.in_top_k(logits, label, 1), tf.float32, name='correct')
        accuracy = tf.reduce_mean(correct, name='accuracy')

        # This will monitor training error (in a moving_average fashion):
        # 1. write the value to tensosrboard
        # 2. write the value to stat.json
        # 3. print the value after each epoch
        train_error = tf.reduce_mean(1 - correct, name='train_error')
        summary.add_moving_summary(train_error, accuracy)

        # Use a regex to find parameters to apply weight decay.
        # Here we apply a weight decay on all W (weight matrix) of all fc layers
        wd_cost = tf.multiply(1e-5,
                              regularize_cost('fc.*/kernel', tf.nn.l2_loss),
                              name='regularize_loss')
        self.cost = tf.add_n([wd_cost, cost], name='total_cost')
        summary.add_moving_summary(cost, wd_cost, self.cost)

        # monitor histogram of all weight (of conv and fc layers) in tensorboard
        summary.add_param_summary(('.*/kernel', ['histogram', 'rms']))

    def build_graph(self, input, nextinput):
        is_training = get_current_tower_context().is_training
        initializer = tf.random_uniform_initializer(-0.05, 0.05)

        def get_basic_cell():
            cell = rnn.BasicLSTMCell(num_units=HIDDEN_SIZE, forget_bias=0.0, reuse=tf.get_variable_scope().reuse)
            if is_training:
                cell = rnn.DropoutWrapper(cell, output_keep_prob=1 - DROPOUT)
            return cell

        cell = rnn.MultiRNNCell([get_basic_cell() for _ in range(NUM_LAYER)])

        def get_v(n):
            return tf.get_variable(n, [BATCH, HIDDEN_SIZE],
                                   trainable=False,
                                   initializer=tf.constant_initializer())

        state_var = [rnn.LSTMStateTuple(
            get_v('c{}'.format(k)), get_v('h{}'.format(k))) for k in range(NUM_LAYER)]
        self.state = state_var = tuple(state_var)

        embeddingW = tf.get_variable('embedding', [VOCAB_SIZE, HIDDEN_SIZE], initializer=initializer)
        input_feature = tf.nn.embedding_lookup(embeddingW, input)  # B x seqlen x hiddensize
        input_feature = Dropout(input_feature, keep_prob=1 - DROPOUT)

        with tf.variable_scope('LSTM', initializer=initializer):
            input_list = tf.unstack(input_feature, num=SEQ_LEN, axis=1)  # seqlen x (Bxhidden)
            outputs, last_state = rnn.static_rnn(cell, input_list, state_var, scope='rnn')

        # update the hidden state after a rnn loop completes
        update_state_ops = []
        for k in range(NUM_LAYER):
            update_state_ops.extend([
                tf.assign(state_var[k].c, last_state[k].c),
                tf.assign(state_var[k].h, last_state[k].h)])

        # seqlen x (Bxrnnsize)
        output = tf.reshape(tf.concat(outputs, 1), [-1, HIDDEN_SIZE])  # (Bxseqlen) x hidden
        logits = FullyConnected('fc', output, VOCAB_SIZE,
                                activation=tf.identity, kernel_initializer=initializer,
                                bias_initializer=initializer)
        xent_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
            logits=logits, labels=tf.reshape(nextinput, [-1]))

        with tf.control_dependencies(update_state_ops):
            cost = tf.truediv(tf.reduce_sum(xent_loss),
                              tf.cast(BATCH, tf.float32), name='cost')  # log-perplexity

        perpl = tf.exp(cost / SEQ_LEN, name='perplexity')
        summary.add_moving_summary(perpl, cost)
        return cost

    def _build_graph(self, inputs):
        image_pos = inputs[0]
        image_pos = image_pos / 128.0 - 1

        z = tf.random_uniform([args.batch, args.z_dim], minval=-1, maxval=1, name='z_train')
        z = tf.placeholder_with_default(z, [None, args.z_dim], name='z')

        def summary_image(name, x):
            x = (x + 1.0) * 128.0
            x = tf.clip_by_value(x, 0, 255)
            tf.summary.image(name, tf.cast(x, tf.uint8), max_outputs=30)

        with argscope([Conv2D, FullyConnected],
                      W_init=tf.truncated_normal_initializer(stddev=0.02)):
            with tf.variable_scope('gen'):
                image_gen = self.decoder(z)

            with tf.variable_scope('discrim'):
                with tf.variable_scope('enc'):
                    hidden_pos = self.encoder(image_pos)
                    hidden_neg = self.encoder(image_gen)

                with tf.variable_scope('dec'):
                    recon_pos = self.decoder(hidden_pos)
                    recon_neg = self.decoder(hidden_neg)

        with tf.name_scope('viz'):
            summary_image('generated-samples', image_gen)
            summary_image('reconstruct-real', recon_pos)
            summary_image('reconstruct-fake', recon_neg)

        with tf.name_scope('losses'):
            L_pos = tf.reduce_mean(tf.abs(recon_pos - image_pos), name='loss_pos')
            L_neg = tf.reduce_mean(tf.abs(recon_neg - image_gen), name='loss_neg')

            eq = tf.subtract(GAMMA * L_pos, L_neg, name='equilibrium')
            measure = tf.add(L_pos, tf.abs(eq), name='measure')

            kt = tf.get_variable('kt', dtype=tf.float32, initializer=0.0)

            update_kt = kt.assign_add(1e-3 * eq)
            with tf.control_dependencies([update_kt]):
                self.d_loss = tf.subtract(L_pos, kt * L_neg, name='loss_D')
                self.g_loss = L_neg

        add_moving_summary(L_pos, L_neg, eq, measure, self.d_loss)
        tf.summary.scalar('kt', kt)

        self.collect_variables()